Caffe Re-implementation of PointNet [Qi et al., CVPR 2017]

This repository contains caffe implementation of PointNet, which is described in:

PointNet: Deep Learning on Point Sets for 3D Classification and Segmentation, CVPR 2017, Oral,
Charles R. Qi*, Hao Su*, Kaichun Mo, and Leonidas J. Guibas (* equal contribution)

Dependencies

• simbaforrest's Caffe that supports Matrix Multiplication Layer, Permute Layer
• unzip

Usage

Installation

Clone the repository recursively:

git clone --recursive https://github.com/xinghaochen/pointnet.caffe.git

Install caffe:

cd caffe
cp Makefile.config.example Makefile.config
# uncomment WITH_PYTHON_LAYER := 1
# change other settings accordingly
make -j16
make pycaffe -j16

Add path/to/pointnet.caffe/libs and path/to/pointnet.caffe/caffe/python to PYTHONPATH

Training

Download and unzip the data:

cd data
sh download_data_modelnet40.sh

For basic classification network without data augmentation:

sh train_pointnet_cls_basic.sh

For basic classification network with data augmentation:

sh train_pointnet_cls_basic_aug.sh

For classification network with input transform (3x3):

sh train_pointnet_cls_input_tnet.sh

For classification network with input transform (3x3) and feature transform (64x64):

sh train_pointnet_cls.sh

Testing

sh test_pointnet_cls_basic.sh

or

sh test_pointnet_cls_basic_aug.sh

or

sh test_pointnet_cls_input_tnet.sh

or

sh test_pointnet_cls.sh

Performance

Classification results on ModelNet40:

Network	Original Paper	Ours w/o aug	Ours w/ aug	model name
PointNet basic cls	87.1%	88.07%	85.39	pointnet_cls_basic
PointNet w/ input T-Net	87.9%	89.53%	-	pointnet_cls_input_tnet
PointNet w/ input & feat T-Net	89.2%	89.41%	-	pointnet_cls

TODOs

• Basic PointNet for Classification
• Basic PointNet for Part/Semantic Segmentation
• Better data augmentation
• T-Nets

Acknowledges

• simbaforrest for his fork of Caffe
• Charles R. Qi for his awesome work of PointNet family.